1 Article information

2
3 Article title
4 Harnessing Deep Learning for Detection of Diabetic Retinopathy in Geriatric Group using Optical Coherence
5 Tomography Angiography-OCTA: A Promising Approach
6
7 Authors
8 Pooja Bidwai a, Shilpa Gite a,*, Biswajeet Pradhanb*,c, Harshita Gupta a, Abdullah Alamrid
9
10 Affiliations
a
11 Symbiosis Centre for Applied Artificial Intelligence (SCAAI) Symbiosis Institute of Technology, Symbiosis International
12 (Deemed University) (SIU), Lavale, Pune 412115 India, [email protected],
13 [email protected], [email protected]
b
14 Symbiosis Institute of Technology, Symbiosis International (Deemed University) (SIU), Lavale, Pune 412115 India
c
15 Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS), School of Civil and Environmental
16 Engineering, University of Technology Sydney, NSW 2007, Australia
d
17 Earth Observation Centre, Institute of Climate Change, Universiti Kebangsaan Malaysia, 43600 UKM, Bangi,
18 Selangor, Malaysia
e
19 Department of Geology and Geophysics, College of Science, King Saud University, Riyadh, Saudi Arabia;
20 [email protected]
21
22 Corresponding author’s email address and Twitter handle
23 [email protected]; [email protected]
24
25 Keywords
26 Diabetic Retinopathy, Geriatric Population, Optical Coherence Tomography Angiography (OCTA), Convolutional
27 Neural Networks, Deep Learning, Classification
28
29 Abstract
30 The prevalence of diabetic retinopathy (DR) among the geriatric population poses significant challenges for early
31 detection and management. Optical Coherence Tomography Angiography (OCTA) combined with Deep Learning
32 presents a promising avenue for improving diagnostic accuracy in this vulnerable demographic. In this method, we
33 propose an innovative approach utilizing OCTA images and Deep Learning algorithms to detect diabetic retinopathy in
34 geriatric patients. We have collected 262 OCTA scans of 179 elderly individuals, both with and without diabetes, and
35 trained a deep-learning model to classify retinopathy severity levels. Convolutional Neural Network (CNN) models:
36 Inception V3, ResNet-50, ResNet50V2, VggNet-16, VggNet-19, DenseNet121, DenseNet201, EfficientNetV2B0, are
37 trained to extract features and further classify them.
38 Here we demonstrate:
39  The potential of OCTA and Deep Learning in enhancing geriatric eye care at the very initial stage.
40  The importance of technological advancements in addressing age-related ocular diseases and providing
41 reliable assistance to clinicians for DR classification.
42  The efficacy of this approach in accurately identifying diabetic retinopathy stages, thereby facilitating timely
43 interventions, and preventing vision loss in the elderly population.
44
45
46
 1 Graphical abstract
2

3
4
5 Specifications table
6
Subject area Computer science

More specific subject area Image Processing.

Name of your method Detection of Diabetic Retinopathy Using Deep Learning

Name and reference of the original

N/A
method
Optical Coherence Tomography Angiography-OCTA Dataset for Detection of
Diabetic Retinopathy [Data set]. Zenodo.
Resource availability
https://doi.org/10.5281/zenodo.10400092

7
8 Background
9 Diabetic retinopathy (DR) is a complex disease involving microvascular damage, inflammation, and neurodegeneration
10 in the retina, highlighting the need for innovative treatments and early detection methods [1]. Diagnostic tools for DR
11 include ophthalmoscopy, fundus photography, fundus fluorescein angiography (FFA), optical coherence tomography
12 (OCT), and optical coherence tomography angiography (OCTA), with FFA being the traditional gold standard [2].
13 However, OCTA offers a non-invasive alternative, utilizing motion contrast imaging to generate volumetric
14 angiography images without the need for dye injection, presenting a safer option for patients [3]. Artificial intelligence
15 (AI) is poised to revolutionize DR screening by enhancing accessibility, cost-effectiveness, and efficiency through
16 telehealth and by improving the accuracy of retinal image analysis for early diagnosis [4],[5],[13]. In the realm of
17 medical imaging and diagnostics, the classification of Diabetic Retinopathy (DR) has garnered significant attention,
18 generating a wealth of research aimed at enhancing detection and grading methodologies. Notably, the application of
19 Convolutional Neural Networks (CNN) has emerged as a cornerstone in this field, offering promising avenues for
20 accurate and efficient diagnosis. Gondal et al. [6] proposed a CNN model for the identification of referable diabetic
21 retinopathy (RDR), leveraging two publicly available fundus image datasets. Their approach involved a binary
22 classification system, distinguishing between non-recommendable and recommendable DR, based on the severity of
23 the condition. Training on the Kaggle dataset and testing on DiaretDB1, they achieved a sensitivity of 93.6% and a
 1 specificity of 97.6%, marking a significant stride in the automated detection of RDR. Building on this, S. Qummar et al.
2 [7] explored the efficacy of an ensemble of five CNN models, including Inceptionv3, Resnet50, Xception, Densenet169,
3 and Densenet121, on a Kaggle dataset comprising retinal images. This ensemble was engineered to harness the rich
4 feature sets inherent in the data, aiming to elevate accuracy across all DR stages. Their findings underscored the
5 potential of their model to outperform existing technologies, demonstrating superior detection capabilities of DR
6 stages, and offering a substantial improvement over state-of-the-art methods utilizing the same dataset. Further
7 innovations are evident in the work of researchers [8], who introduced a three-step program employing octal
8 coherence tomography (OCT) for DR detection. This method focuses on segmenting retinal layers within OCT images,
9 extracting 3-dimensional features that reflect both the first-order reflectivity and the thickness of the layers. Utilizing
10 backpropagation neural networks for classification, they achieved an accuracy of 96.81% through leave-one-subject-
11 out cross-validation, confirming the superiority of their method. In addition, the development of an automated
12 Diabetic Retinopathy staging system using CNN for analyzing OCT images [9] marks a significant advance. This system
13 not only outperformed traditional machine learning models but also surpassed human experts in accuracy and
14 reliability for DR detection, achieving a quadratic weighted κ of 0.908 for the six-stage leveling task. Moreover, the use
15 of a Multiple Instance Learning (MIL)-based CNN, specifically MIL-ResNet14, has shown promise for classifying DR in a
16 weakly labeled dataset of widefield OCTA enface retina images [10]. This method demonstrated a remarkable
17 resilience against adversarial attacks and high classification scores, highlighting its practical utility in clinical settings.
18 The exploration of transfer learning for classifying OCTA images from a small dataset [11] and its application in a CNN
19 model with the VGG16 architecture for DR detection using OCTA images [12] further illustrate the evolving landscape
20 of DR diagnostics. These studies emphasize the potential of geometric data augmentation and transfer learning to
21 enhance accuracy and model performance across various imaging devices, addressing critical research gaps such as
22 the optimization of preprocessing techniques, the need for longitudinal studies, the interpretability of deep learning
23 models, and tackling data imbalance and biases.
24
25 The development of this methodology for detecting diabetic retinopathy through Deep Learning applied to Optical
26 Coherence Tomography Angiography (OCTA) images represents a significant advancement in the early diagnosis and
27 management of this diabetes complication, a leading cause of blindness globally. It underscores a pivotal shift towards
28 more efficient, scalable, and objective diagnostic practices, with the potential to significantly improve patient
29 outcomes, reduce healthcare expenditures, and contribute to a deeper understanding of the pathophysiology
30 of diabetic retinopathy.
31
32 Method details
33 The methodology comprises a novel dataset obtained through a case-control study conducted among a geriatric
34 population. It encompasses computational prerequisites, data pre-processing procedures, workflow delineation, CNN
35 model specifications, feature extraction methodologies, and classification tasks. Each aspect is elaborated upon
36 comprehensively in the following sections.

37 Data Description

38 Datasets for Diabetic Retinopathy detection pose various challenges since they require expert guidance and more
39 patient cooperation to obtain optimal clarity of the image [15, 16]. The primary dataset used in this study is the Optical
40 Coherence Tomography Angiography (OCTA) dataset specifically developed for the detection of Diabetic Retinopathy.
41 This dataset was carefully collected at Natasha Eye Care and Research Institute and reflects a concerted effort to
 1 harness advanced imaging technologies for ophthalmic research [17]. The dataset includes 262 high-resolution OCTA
2 images representing a variety of retinal conditions across different stages of diabetic retinopathy.

3 Characteristics of the Dataset

4 Data Source Location: Natasha Eye Care and Research Institute, Pimple Saudagar, Pune.

5 Acquisition: Nonmydriatic OCTA images were acquired using the Optovue Avanti Edition machine. Images are classified
6 into three categories: NO DR signs, Mild DR, and Moderate DR.

7 Type of Data: High-resolution OCTA images, offering detailed views of the retinal vasculature, which are crucial for
8 accurate detection and classification of diabetic retinopathy stages. OCTA (8x8 mm) images with dimensions 1596 ×
9 990, 96 dpi, and jpeg. The dataset contains 262 OCTA images, carefully annotated, and verified by retina experts. Fig
10 1. Shows sample OCTA images from the dataset.

11

12 Figure 1. Sample OCTA images from the dataset.

13 Computational Requirements:

14 Hardware: Adequate computational resources are crucial for handling the processing demands of deep learning
15 models and large datasets efficiently. A multi-core CPU is necessary for basic computation, while a GPU is highly
16 recommended to significantly speed up the training and feature extraction processes involving deep learning models.

17 Memory: Sufficient RAM (at least 16GB, with 32GB or more being ideal) is required to manage large datasets and the
18 operations of complex models in memory without significant slowdowns.

19 Software: The notebook is run in a Jupyter environment, which requires installation via Anaconda or directly through
20 Python pip. Dependencies include TensorFlow, scikit-learn, pandas, and NumPy, among others.

21 Data Preprocessing

22 The data preparation phase is crucial for conditioning the images for optimal analysis by the CNN architectures. The
23 raw OCTA images obtained, such as the one exemplified in Figure 1 contain several layers of retinal structure that are
24 relevant for the diagnosis and evaluation of Diabetic Retinopathy. Each of these layers offers unique information
25 essential for the comprehensive analysis of retinal health. To prepare the images for feature extraction, the following
26 preprocessing steps were applied to each image in the dataset:
 1 Layer Separation: Initially, OCTA images are composed of different retinal layers, including the Angio-Superficial,
2 Angio-Deep, Angio-Outer Retina, and Angio-Choriocapillaris, as well as cross-sectional optical coherence tomography
3 angiography (OCTA) scans. The first step in preprocessing was to separate these layers. Each important layer was
4 cropped from the raw image using a region of interest (ROI) approach. This step was performed carefully to ensure
5 the capture of the entire vascular network and relevant anatomical features without introducing artifacts.

6 Resizing: Once the layers were separated, each cropped image segment was resized to maintain a consistent input
7 dimension across all images for the chosen CNN architectures. This standardization is vital to ensure that each model
8 receives input of the same scale, which is crucial for comparative analysis.

9 Normalization: Following resizing, the pixel values in each cropped layer were normalized. This normalization is not
10 merely a scale adjustment to the range of 0 to 1 but also accounts for variations in imaging brightness and contrast
11 that may occur due to different patient conditions or imaging equipment. The purpose of this step is to reduce model
12 sensitivity to these inconsistencies, which are unrelated to the disease markers of interest.

13 Augmentation: To address the limited size of the dataset and potential overfitting, data augmentation techniques
14 were implemented. These included geometric transformations such as rotations and horizontal flips, as well as
15 brightness and contrast adjustments. The augmentation was carefully calibrated to reflect realistic variations in OCTA
16 imaging, enhancing the robustness of the subsequent feature extraction process.

17 Quality Assessment: Each image was subject to a quality assessment, as OCTA image quality can vary significantly due
18 to patient movement or other factors during acquisition. Only images with a quality score above a predetermined
19 threshold were included in the study to ensure reliable feature extraction. Images like the one provided, with a scan
20 quality of 7/10, required careful consideration to determine if they met the quality criteria for inclusion.

21 Final Review: The processed images underwent a final review to ensure that all preprocessing steps were executed
22 consistently and that the resulting images were ready for feature extraction. This review was an integral quality control
23 step to prevent the introduction of biases in the comparative analysis of the CNN architectures.

24 By meticulously executing these preprocessing steps, we can ensure that the extracted features from each CNN
25 architecture are truly representative of the underlying pathology and are not influenced by imaging artifacts or
26 inconsistencies. This meticulous approach to data preparation is expected to yield more reliable and comparable
27 results in our study.

28 Workflow:

29 Import Necessary Libraries:

30 Import libraries necessary for data manipulation, image processing, machine learning, and deep learning tasks. Key
31 libraries include NumPy for numerical operations, pandas for data manipulation, TensorFlow for constructing and
32 training deep learning models, and sci-kit-learn for various machine learning utilities.

33 Setup CNN as Feature Extractor:

 1 Initialize a Convolutional Neural Network (CNN) model pre-trained on ImageNet data, such as VGG19. Configure the
2 model to use its penultimate layer ('fc2') as a feature extractor. This layer is chosen because it provides a robust and
3 comprehensive representation of the image data, suitable for feature extraction.

4 Data Loading and Feature Extraction:

5 Load image metadata from a CSV file that contains paths to images and their corresponding labels. Preprocess the
6 images to fit the input requirements of the pre-trained CNN model. This involves resizing images to the required
7 dimensions, converting them into arrays, and normalizing the pixel values to match the preprocessing used in the CNN
8 training.
9 Pass the preprocessed images through the CNN to extract features. These features are then compiled into a dataset
10 that is ready for use with machine learning models.

11 Data Splitting:

12 Divide the dataset containing the extracted features and their corresponding labels into training and testing sets using
13 a standard train-test split approach. The split is made in an 80%-20% ratio, with 80% of the data used for training the
14 models and 20% reserved for testing and model evaluation. This split is crucial for training the models on a large subset
15 of data while still retaining a separate subset for an unbiased evaluation of model performance.

16 Classifier Training and Evaluation:

17 Train two types of classifiers on the extracted features: a K-Nearest Neighbors (KNN) classifier and a simple feed-
18 forward Neural Network. These models are chosen for their differing approaches to classification tasks, with KNN
19 being a simple, distance-based algorithm and the Neural Network providing a more complex, nonlinear decision
20 boundary.
21 Evaluate each classifier on the test set. Key performance metrics such as accuracy, precision, recall, and F1-score are
22 calculated to assess each model's ability to accurately classify new, unseen images. This evaluation helps determine
23 the effectiveness of the feature extraction and the overall utility of the trained classifiers.

24 Overview of CNN Architectures

25 ResNet50V2: In our study, we employed the ResNet50V2 architecture, known for its advanced residual learning
26 framework with 50 layers featuring shortcut connections that prevent the vanishing gradient problem. Initially trained
27 on the comprehensive ImageNet dataset, ResNet50V2 was adapted to diagnose diabetic retinopathy using OCTA
28 retinal images. This adaptation involved using the 'avg_pool' layer for extracting essential global contextual
29 information from the images, critical for identifying subtle indicators of the disease such as microvascular changes and
30 hemorrhages. The model's capability to handle deep networks efficiently through its residual blocks enabled it to learn
31 discriminative features necessary for medical diagnosis.
 1

2 Figure 2. RESNET50V2

3 EfficientNetV2B0: In our study, we utilized EfficientNetV2B0, a model renowned for its scalability and efficiency in
4 handling a range of different image resolutions, adapting seamlessly to varying computational constraints. Initially
5 trained on the extensive ImageNet dataset, EfficientNetV2B0 was adapted to specifically address the complex patterns
6 characteristic of diabetic retinopathy in OCTA retinal images. This model is distinguished by its compound scaling
7 method, which uniformly scales the network width, depth, and resolution with a set of fixed scaling coefficients [24].

8 The EfficientNetV2B0 model was particularly effective due to its ability to process images at different resolutions while
9 maintaining model efficiency.

10 DenseNet121: In our study, DenseNet121 was employed due to its distinctive architectural features that contribute to
11 its efficiency in training and accuracy in performance. Known for its dense connectivity pattern [26], DenseNet121
12 connects each layer to every other layer in a feed-forward fashion, which significantly improves the network's ability
13 to reuse features and reduces the problem of vanishing gradients. Initially trained on the ImageNet dataset,
14 DenseNet121 was adapted for the specific task of classifying diabetic retinopathy using OCTA retinal images. This
15 adaptation involved leveraging the model's feature-reusing capability to enhance the extraction of complex patterns
16 indicative of the disease, such as microvascular changes and hemorrhages. The use of DenseNet121 in our diagnostic
17 approach emphasizes its potential to improve the precision and reliability of medical imaging analyses, making it a
18 valuable tool in clinical settings where diagnostic accuracy is crucial.
 1

2 Figure 3. DenseNet121

3 DenseNet201: In our research, we implemented DenseNet201, a model celebrated for its unique architectural
4 features that facilitate efficient feature utilization and gradient flow. DenseNet201, an extension of the DenseNet
5 architecture, is characterized by its deeper structure and more layers, allowing for even more complex feature learning
6 and reuse across the network. DenseNet201's architecture employs a dense connectivity pattern where each layer is
7 directly connected to every subsequent layer, ensuring maximum information flow between layers. This setup
8 significantly reduces the vanishing gradient problem and enhances feature propagation, which is crucial for capturing
9 the subtle nuances in medical images indicative of diabetic retinopathy, such as delicate vascular structures and early
10 signs of hemorrhages.

11
 1

2 Figure 4. DenseNet201

3 Inception V3: It is a sophisticated convolutional neural network (CNN) that exemplifies innovation in network
4 architecture design [25]. It is the third iteration of the inception family, characterized by its inception modules. These
5 modules are smaller convolutions inside a larger convolution block, enabling the network to choose from various filter
6 sizes (1x1, 3x3, 5x5) within the same layer. The inclusion of these parallel paths optimizes computational efficiency by
7 allowing the model to adapt to the spatial hierarchies of features in images. This modular approach reduces the
8 number of parameters compared to a densely connected network, minimizing computational load without
9 compromising depth or breadth. For Diabetic Retinopathy detection, the varied granularity of feature extraction
10 offered by Inception V3 can be particularly advantageous for capturing the intricate vascular changes characteristic of

11 the disease.

12 Figure 5. Inception V3 architecture

 1 VGG16: VGG16 is celebrated for its simplicity and uniformity, consisting of a series of convolutional layers followed
2 by max-pooling layers, and fully connected layers towards the end [23]. The architecture uses small 3x3 receptive
3 fields throughout the entire network, maintaining a fixed convolution stride, which is effective in capturing image
4 features. The depth of the network, with 16 weight layers, allows it to learn from a large amount of data. However,
5 this simplicity also results in a considerable number of parameters, which may lead to intensive computational
6 demands. In the context of OCTA images, VGG16's ability to capture fine-grained details through its uniform
7 convolutions could be particularly useful for highlighting features such as microaneurysms or hemorrhages that are
8 critical for accurate diagnosis.

10 Figure 6. Representation of VGG16 architecture

11

12 Feature Extraction Process

13 The pre-trained models used, including VGG16, Inception V3, ResNet50 V2, EfficientNetV2 B0, DenseNet121, and
14 DenseNet201, are equipped with weights from extensive training on the ImageNet dataset, which captures a broad
15 spectrum of visual concepts and features. For efficient feature extraction from OCTA images, the classification layers
16 of these networks were removed, allowing the core architectures to output detailed feature vectors. Specifically,
17 features were extracted from the global average pooling layers in Inception V3 and ResNet50 V2 to reduce spatial
18 dimensions while preserving important spatial hierarchies. In contrast, VGG19 utilized the last max-pooling layer,
19 resulting in larger feature vectors due to the network's depth. EfficientNetV2 B0 adapted its depth for optimal feature
20 scaling, and DenseNet121 and DenseNet201 leveraged their dense connectivity before transition layers to maintain
21 information flow. This approach ensures that each model captures the most informative, higher-level abstract features
22 necessary for accurately classifying stages of Diabetic Retinopathy, thus optimizing the discriminative capabilities of
23 the extracted features.

24 Classification Tasks

25 The extracted features were fed into various classification algorithms. The choice of classifiers was guided by the need
26 to evaluate the quality of the features in terms of their ability to distinguish between different classes in the OCTA
27 dataset. Two main types of classifiers used are K-nearest neighbors (KNN) and Neural Networks.

28 KNN is a simple, non-parametric, and versatile algorithm used for classification and regression. It was chosen
29 for its efficacy in pattern recognition which can be particularly beneficial when dealing with medical images where the
 1 distinction between classes may be subtle. The ease of understanding and implementing KNN, along with its capacity
2 to make strong predictions with sufficient training data, made it a suitable choice for this study.

3 Neural Networks were used due to their ability to learn non-linear relationships and their adaptability in handling
4 complex patterns in data. Their deep learning capabilities allow for a more nuanced understanding of medical images,
5 which is essential for tasks such as identifying stages of Diabetic Retinopathy where features may not be linearly
6 separable.

7 Method validation
8 The performance of the classifiers, using the features extracted by the CNN architectures, was evaluated using a
9 comprehensive set of metrics:
10 ● Accuracy- Accuracy represents the proportion of true results (both true positives and true negatives) among
11 the total number of cases examined [18]. High accuracy is essential in medical diagnostics to ensure reliable
12 predictions for both the presence and absence of disease.

13 ● Precision- Precision, or positive predictive value, measures the proportion of correctly identified positives
14 out of all positive predictions made [19]. In the context of Diabetic Retinopathy, high precision would
15 indicate that a high percentage of patients identified as having the condition truly have it.

16 ● Recall (Sensitivity)- Recall, also known as sensitivity in medical diagnostics, measures the proportion of
17 actual positives correctly identified [19]. It is crucial to ensure that the model identifies as many patients
18 with the condition as possible.

19 ● F1 Score- The F1 Score is the harmonic mean of precision and recall [19]. It is used when the balance
20 between precision and recall is vital, which is often the case in medical diagnostics, as both false positives
21 and false negatives carry significant consequences.

22 ● AUC Score- The Area Under the Receiver Operating Characteristic curve (AUC-ROC) is a performance
23 measurement for the classification tasks at various threshold settings[20]. The AUC Score provides an
24 aggregate measure of performance across all possible classification thresholds.

25 ● Kappa's Coefficient-Kappa’s Coefficient measures inter-rater reliability for qualitative (categorical) items. It is
26 used to assess the agreement between two raters on the assignment of categories [21]. In this study, it
27 gauges the consistency of the classification results against a standard.

28 ● Specificity- Specificity measures the proportion of actual negatives correctly identified (true negative rate)
29 [22]. For Diabetic Retinopathy, high specificity means that healthy individuals are correctly identified as not
30 having the disease, which is critical for avoiding unnecessary medical interventions.
31
32 Each metric offers a unique insight into the classifiers' performance and collectively provides a comprehensive
33 evaluation of the models' effectiveness. The balance of these metrics is especially pertinent in medical
34 applications, where the cost of false positives or negatives can be high, and the classifier's ability to generalize
35 well and provide reliable diagnostics is paramount. By leveraging this suite of evaluation metrics, the study aims
36 to present a detailed analysis of the classifiers' performance in Diabetic Retinopathy detection, offering a nuanced
37 understanding of the strengths and weaknesses of the feature extraction capabilities of each CNN architecture.
38
1 Table 1. Final Validated Results
2
3
VGG19
Accuracy Accuracy Accuracy
Model Precision Recall F-score AUC Kappa N0_DR Mild_DR Moderate_DR
Logistic Regression 0.95 0.95 0.95 0.99 0.91 0.95 0.92 1
SVM 0.89 0.88 0.88 0.9 0.78 0.98 0.8 0.63
Decision Tree 0.66 0.64 0.65 0.71 0.39 0.67 0.6 0.62
Random Forest 0.88 0.87 0.86 0.97 0.75 0.98 0.8 0.5
Gradient Boosting 0.87 0.87 0.86 0.94 0.76 0.91 0.88 0.62
KNN 0.93 0.91 0.91 0.99 0.84 0.86 0.96 1
Neural Network 0.31 0.56 0.4 0.5 0.4 1 0.91 0.88

ResNet50V2
Accuracy Accuracy Accuracy
Model Precision Recall F-score AUC Kappa N0_DR Mild_DR Moderate_DR
Logistic Regression 0.87 0.87 0.87 0.97 0.77 0.86 0.84 1
SVM 0.84 0.83 0.82 0.86 0.69 0.86 0.88 0.5
Decision Tree 0.65 0.64 0.64 0.68 0.37 0.69 0.64 0.38
Random Forest 0.84 0.83 0.81 0.94 0.68 0.93 0.84 0.25
Gradient Boosting 0.9 0.89 0.89 0.95 0.8 0.95 0.88 0.63
KNN 0.94 0.93 0.93 0.99 0.88 0.93 0.96 0.88
Neural Network 0.92 0.92 0.92 0.98 0.86 0.93 0.88 1

EfficientNetV2B0
Accuracy Accuracy Accuracy
Model Precision Recall F-score AUC Kappa N0_DR Mild_DR Moderate_DR
Logistic Regression 0.92 0.92 0.92 0.99 0.86 0.91 0.92 1
SVM 0.9 0.89 0.89 0.9 0.81 0.91 0.92 0.75
Decision Tree 0.64 0.63 0.63 0.66 0.35 0.67 0.68 0.25
Random Forest 0.91 0.91 0.91 0.99 0.84 0.88 0.96 0.87
Gradient Boosting 0.93 0.93 0.93 0.98 0.88 0.93 0.96 0.87
KNN 0.9 0.84 0.85 1 0.74 0.71 1 1
Neural Network 0.92 0.92 0.92 0.99 0.86 0.91 0.92 1
 DenseNet121
Accuracy Accuracy Accuracy
Model Precision Recall F-score AUC Kappa N0_DR Mild_DR Moderate_DR
Logistic Regression 0.95 0.95 0.95 0.99 0.91 0.95 0.92 1
SVM 0.9 0.89 0.89 0.89 0.8 0.98 0.88 0.5
Decision Tree 0.62 0.57 0.58 0.69 0.31 0.6 0.48 0.75
Random Forest 0.9 0.89 0.89 0.99 0.81 0.93 0.88 0.75
Gradient Boosting 0.92 0.89 0.9 0.99 0.82 0.81 1 1
KNN 0.94 0.92 0.92 1 0.86 0.86 1 1
Neural Network 0.91 0.91 0.91 0.99 0.83 0.91 0.92 0.88

DenseNet201
Accuracy Accuracy Accuracy
Model Precision Recall F-score AUC Kappa N0_DR Mild_DR Moderate_DR
Logistic Regression 0.96 0.96 0.96 1 0.93 0.98 0.92 1
SVM 0.93 0.92 0.92 0.92 0.85 1 0.88 0.63
Decision Tree 0.68 0.64 0.65 0.69 0.39 0.64 0.72 0.38
Random Forest 0.91 0.91 0.9 0.98 0.83 0.98 0.88 0.63
Gradient Boosting 0.95 0.95 0.95 0.99 0.9 0.98 0.92 0.88
KNN 0.9 0.89 0.89 0.98 0.82 0.86 0.92 1
Neural Network 0.98 0.97 0.97 1 0.95 1 0.92 1

InceptionV3
Accuracy Accuracy Accuracy
Model Precision Recall F-score AUC Kappa N0_DR Mild_DR Moderate_DR
Logistic Regression 0.92 0.91 0.91 1 0.84 0.86 0.96 1
SVM 0.89 0.88 0.88 0.88 0.78 0.91 0.92 0.63
Decision Tree 0.73 0.68 0.7 0.72 0.47 0.71 0.72 0.38
Random Forest 0.91 0.91 0.9 0.98 0.83 0.93 0.96 0.63
Gradient Boosting 0.89 0.88 0.88 0.98 0.79 0.86 0.96 0.75
KNN 0.91 0.88 0.88 0.99 0.8 0.79 1 1
Neural Network 0.93 0.92 0.92 0.99 0.86 0.88 0.96 1

VGG16
Accuracy Accuracy Accuracy
Model Precision Recall F-score AUC Kappa N0_DR Mild_DR Moderate_DR
Logistic Regression 0.91 0.91 0.91 0.99 0.84 0.91 0.92 0.88
SVM 0.9 0.89 0.89 0.89 0.8 0.98 0.88 0.5
Decision Tree 0.76 0.71 0.72 0.78 0.52 0.67 0.8 0.63
Random Forest 0.87 0.85 0.84 0.98 0.72 0.95 0.84 0.38
Gradient Boosting 0.9 0.89 0.89 0.98 0.81 0.93 0.88 0.75
KNN 0.95 0.95 0.95 0.99 0.91 0.95 0.96 0.88
Neural Network 0.88 0.88 0.88 0.99 0.79 0.88 0.88 0.88
 ResNet50
Accuracy Accuracy Accuracy
Model Precision Recall F-score AUC Kappa N0_DR Mild_DR Moderate_DR
Logistic Regression 0.95 0.95 0.95 0.98 0.91 0.95 0.92 1
SVM 0.89 0.88 0.88 0.87 0.78 0.95 0.84 0.63
Decision Tree 0.76 0.75 0.75 0.77 0.56 0.81 0.72 0.5
Random Forest 0.9 0.89 0.89 0.96 0.8 1 0.8 0.63
Gradient Boosting 0.87 0.87 0.86 0.97 0.75 0.93 0.84 0.63
KNN 0.93 0.92 0.92 0.99 0.86 0.88 0.96 1
Neural Network 0.91 0.91 0.91 0.99 0.83 0.93 0.88 0.88
1
2 Discussion:

3 In a comprehensive analysis comparing classifiers like Logistic Regression, Support Vector Machine,
4 Decision Trees, Random Forest, Gradient Boosting, K-Nearest Neighbors (KNN) and Neural Networks (NN)
5 across eight different architectures like ResNet50, ResNet50V2, EfficientNetV2B0, DenseNet121,
6 DenseNet201, InceptionV3, VGG16, and VGG19 varied performance metrics were observed that highlight
7 the strengths and weaknesses of each model depending on the architecture. KNN consistently demonstrated
8 robust performance with high scores across the board, particularly in accuracy, precision, recall, and F-score
9 metrics. For example, in the VGG16 architecture, KNN achieved its highest metrics with a precision of 0.95,
10 recall of 0.95, and an F-score of 0.95, along with an AUC of 0.98, showcasing its efficacy in accurately
11 classifying images with high consistency.

12 On the other hand, Neural Networks showed a strong ability to separate classes effectively, as evidenced by
13 consistently high AUC scores across all architectures. The NN model excelled particularly with the
14 DenseNet201 architecture, achieving the highest weighted precision, recall, and F-score values of 0.97, 0.97,
15 and 0.97, respectively, and an impressive AUC of 0.99. This indicates that Neural Networks can leverage the
16 depth and complexity of certain architectures to achieve superior predictive accuracy and reliability.

17 The results also reveal some challenges. For instance, while KNN performed exceptionally well in terms of
18 overall accuracy and reliability, its performance in EfficientNetV2B0 was relatively lower, with precision,
19 recall, and F-scores around 0.9, 0.84, and 0.85, respectively. This suggests a potential mismatch between the
20 model capabilities and the architectural features of EfficientNetV2B0. Similarly, Neural Networks, despite
21 their high separability capabilities, showed varied effectiveness across different architectures, with lower
22 performance metrics in some cases, like with VGG16 where it recorded the lowest precision, recall, and F-
23 score around 0.88, 0.88, and 0.88, respectively.
 1 This analysis underscores the importance of selecting the appropriate model and architecture combination
2 based on the specific requirements and challenges of the task at hand. While KNN is generally more consistent
3 and reliable for broad use cases, Neural Networks are particularly adept at handling complex and nuanced
4 patterns, making them suitable for tasks that benefit from deep learning's robust feature extraction and
5 classification capabilities.

6
7 Limitations
8 Not applicable
9
10 Ethics statements

11 The study was approved by the Symbiosis Institutional Ethics Committee (BHR) SIU with approval number SIU/IEC/583.
12 Every patient provided informed consent for both treatment and examination taken by the Department of
13 Ophthalmology at Natasha Eye Care and Research Institute. A data usage agreement protects data.

14 CRediT author statement

15 Pooja Bidwai: Conceptualization, Methodology, Data curation, Writing- Reviewing and Editing; Shilpa Gite:
16 Visualization, Investigation, Supervision; Biswajeet Pradhan: Visualization, Resources, Writing- Reviewing and Editing,
17 Funding; Harshita Gupta: Validity tests, Software, Writing- Original draft preparation; Abdullah Alamri: Visualization,
18 Writing- Reviewing and Editing, Funding.

19 Acknowledgments
20
21 This research was funded by the Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS),
22 Faculty of Engineering and IT, University of Technology Sydney and Researchers Supporting Project, King Saud
23 University, Riyadh, Saudi Arabia, under Project RSP2024 R14.
24
25 Declaration of interests
26
27 ☒ The authors declare that they have no known competing financial interests or personal relationships that could
28 have appeared to influence the work reported in this paper.
29
30 ☐ The authors declare the following financial interests/personal relationships which may be considered as potential
31 competing interests:
32
33 Please declare any financial interests/personal relationships which may be considered as potential competing interests
34 here.
35
36 References
 1 [1] Wang, Wei, and Amy C. Y. Lo. 2018. "Diabetic Retinopathy: Pathophysiology and Treatments" International
2 Journal of Molecular Sciences 19, no. 6: 1816. https://doi.org/10.3390/ijms19061816

3 [2] de Barros Garcia JMB, Isaac DLC, Avila M. Diabetic retinopathy, and OCT angiography: clinical findings and future
4 perspectives. Int J Retina Vitreous 2017; 3: 14. DOI: 10.1186/s40942-017-0062-2. PMID: 28293432. PMCID:
5 PMC5346852

6 [3]de Carlo, T.E., Romano, A., Waheed, N.K. et al. A review of optical coherence tomography angiography (OCTA). Int
7 J Retin Vitr 1, 5 (2015). https://doi.org/10.1186/s40942-015-0005-8

8 [4] Gunasekeran, Dinesh V.a,b; Ting, Daniel S.W.a,c; Tan, Gavin S.W.a,c; Wong, Tien Y.a,c. Artificial intelligence for
9 diabetic retinopathy screening, prediction and management. Current Opinion in Ophthalmology 31(5):p 357-365,
10 September 2020. | DOI: 10.1097/ICU.0000000000000693

11 [5] Stevenson, C., Hong, S., & Ogbuehi, K. (2018). Development of an artificial intelligence system to classify
12 pathology and clinical features on retinal fundus images. Clinical & Experimental Ophthalmology, 47, 484 - 489.
13 https://doi.org/10.1111/ceo.13433.

14 [6]W. M. Gondal, J. M. Köhler, R. Grzeszick, G. A. Fink and M. Hirsch, "Weakly-supervised localization of diabetic
15 retinopathy lesions in retinal fundus images," 2017 IEEE International Conference on Image Processing (ICIP), Beijing,
16 China, 2017, pp. 2069-2073, doi: 10.1109/ICIP.2017.8296646.

17 [7]S. Qummar et al., "A Deep Learning Ensemble Approach for Diabetic Retinopathy Detection," in IEEE Access, vol.
18 7, pp. 150530-150539, 2019, doi: 10.1109/ACCESS.2019.2947484.

19 [8] Elgafi M, Sharafeldeen A, Elnakib A, Elgarayhi A, Alghamdi NS, Sallah M, El-Baz A. Detection of Diabetic
20 Retinopathy Using Extracted 3D Features from OCT Images. Sensors (Basel). 2022 Oct 15;22(20):7833. doi:
21 10.3390/s22207833. PMID: 36298186; PMCID: PMC9610651.

22 [9] Translational Vision Science & Technology February 2022, Vol.11, 39. doi:https://doi.org/10.1167/tvst.11.2.39

23 [10] Matten, P., Scherer, J., Schlegl, T. et al. Multiple instance learning based classification of diabetic retinopathy in
24 weakly-labeled widefield OCTA en face images. Sci Rep 13, 8713 (2023). https://doi.org/10.1038/s41598-023-35713-
25 4

26 [11] Andreeva, R, Fontanella, A, Giarratano, Y & Bernabeu, MO 2020, DR Detection Using Optical Coherence
27 Tomography Angiography (OCTA): A Transfer Learning Approach with Robustness Analysis. in International
28 Workshop on Ophthalmic Medical Image Analysis., Chapter 2, Ophthalmic Medical Image Analysis, vol. 12069,
29 Springer, Cham, pp. 11-20. https://doi.org/10.1007/978-3-030-63419-3_2

30 [12] Translational Vision Science & Technology July 2020, Vol.9, 35. doi:https://doi.org/10.1167/tvst.9.2.35

31 [13] Bidwai P, Gite S, Pahuja K, Kotecha K. A Systematic Literature Review on Diabetic Retinopathy Using an Artificial
32 Intelligence Approach. Big Data and Cognitive Computing. 2022; 6(4):152. https://doi.org/10.3390/bdcc6040152
33 Scopus Q1
 1 [14] J. Deng, W. Dong, R. Socher, L. -J. Li, Kai Li and Li Fei-Fei, "ImageNet: A large-scale hierarchical image database,"
2 2009 IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 2009, pp. 248-255, doi:
3 10.1109/CVPR.2009.5206848.

4 [15] Pooja Bidwai. (2023). Multimodal OCTA and Fundus Image dataset for detection of Diabetic Retinopathy
5 (Version-1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.8375220

6 [16] Pooja Bidwai, Shilpa Gite, Aditi Gupta, Kishore Pahuja, Ketan Kotecha, Multimodal dataset using OCTA and
7 fundus images for the study of diabetic retinopathy, Data in Brief, 2024, 110033, ISSN 2352-3409,
8 https://doi.org/10.1016/j.dib.2024.110033 Scopus Q1

9 [17] 6. Bidwai, P., Gite, S., Gupta, A., & Pahuja, K. (2023). Optical Coherence Tomography Angiography-OCTA
10 Dataset for Detection of Diabetic Retinopathy [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10400092

11 [18] Jin Huang and C. X. Ling, "Using AUC and accuracy in evaluating learning algorithms," in IEEE Transactions on
12 Knowledge and Data Engineering, vol. 17, no. 3, pp. 299-310, March 2005, doi: 10.1109/TKDE.2005.50.

13 [19] D. M. Powers, “Evaluation: from precision, recall and f-measure to roc, informedness, markedness and
14 correlation,” Journal of Machine Learning Technologies, vol. 2, no. 1, pp. 37–63, 2011.

15 [20] Andrew P. Bradley, The use of the area under the ROC curve in the evaluation of machine learning algorithms,
16 Pattern Recognition, Volume 30, Issue 7, 1997, Pages 1145-1159, ISSN 0031-3203, https://doi.org/10.1016/S0031-
17 3203(96)00142-2.

18 [21] Julius Sim, Chris C Wright, The Kappa Statistic in Reliability Studies: Use, Interpretation, and Sample Size
19 Requirements, Physical Therapy, Volume 85, Issue 3, 1 March 2005, Pages 257–268,
20 https://doi.org/10.1093/ptj/85.3.257

21 [22] Zhu, W.; Zeng, N.; Wang, N. Sensitivity, specificity, accuracy, associated confidence interval and ROC analysis
22 with practical SAS implementations. In Proceedings of the NESUG Proceedings: Health Care and Life Sciences,
23 Baltimore, MD, USA, 14 November 2010; Volume 19, p. 67.

24 [23] Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional networks for large-scale image recognition."
25 arXiv preprint arXiv:1409.1556 (2014).

26 [24] Tan, Mingxing & Le, Quoc. (2021). EfficientNetV2: Smaller Models and Faster Training.

27 [25] Szegedy, Christian, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. "Rethinking the inception
28 architecture for computer vision." In Proceedings of the IEEE conference on computer vision and pattern
29 recognition, pp. 2818-2826. 2016.

30 [26] Huang, Gao & Liu, Zhuang & van der Maaten, Laurens & Weinberger, Kilian. (2017). Densely Connected
31 Convolutional Networks. 10.1109/CVPR.2017.243.

32 [27] P Bidwai, S Khairnar, S Gite - Explainable AI in Healthcare, Chapter- Explainable Artificial Intelligence in Breast
33 Cancer Identification in Taylor and Francis Group, Page-148-165, Publisher Chapman and Hall/CRC,
34 https://www.taylorfrancis.com/chapters/edit/10.1201/9781003333425-8/explainable-artificial-intelligence-breast-
35 cancer-identification-pooja-bidwai-smita-khairnar-shilpa-gite
1 [28] P. Bidwai, S. Gite, K. Patwa, K. Maheshwari, T. S. Bais and K. Batavia, "Detection of Diabetic Retinopathy using
2 Deep Learning," 2023 IEEE 8th International Conference for Convergence in Technology (I2CT), Lonavala, India, 2023,
3 pp. 1-8, doi: 10.1109/I2CT57861.2023.10126384.